papr-review

name	papr-review
description	Sends a paper for external blind review via the Codex plugin (latest GPT model) and compares external findings with the internal panel. Use when the user says "external review", "get another opinion", "blind review", or after papr-write completes.
argument-hint	[paper_dir]
allowed-tools	["Read","Write","Edit","Grep","Glob","Bash","Agent"]

External Blind Review (via Codex plugin)

Sends finished paper to Codex for independent blind review using its current default GPT model (GPT-5.4 as of April 2026 -- Codex auto-routes, do not hardcode the version).

Prerequisites

Install the Codex plugin:

/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/codex:setup

If the plugin is not installed: write External review skipped -- Codex plugin not installed. to papr-runs/latest-run/latest/ROUND_STATE.md and return.

How to send the review

Use the codex:codex-cli-runtime skill to send the prompt to Codex. If that skill is unavailable, log the failure and skip rather than falling back to any other path -- a contaminated review (e.g., one that picked up internal panel context) is worse than no external review.

On invocation

CRITICAL: Do NOT read ROUND_STATE.md before sending to Codex. Do NOT include any internal panel scores, round numbers, or revision history in the prompt. Do NOT explore the .claude/ or papr-runs/ directories or look for previous run history. Do NOT run ls .claude/, ls papr-runs/, find .claude/, find papr-runs/, or any filesystem exploration there. Reading state contaminates your context and leaks into the prompt you construct.

1. Read paper .tex files from [paper_dir]. Strip \author{} commands.
2. Construct the review prompt below with ONLY the paper text.
3. Send to Codex.
4. AFTER receiving the review, read ROUND_STATE.md (current run only) to write the comparison.

Isolation

External reviewer receives ONLY paper text. Never include: round number, score history, change log, internal discussion, revision framing, or any indication this paper has been reviewed or revised before.

Review Prompt

You are the most critical reviewer on the program committee of a top-tier
ML/CV conference (NeurIPS, ICLR, CVPR level).

Be critical and cautious in your evaluation. If a paper is bad or you are
unsure about its quality, give it bad scores and recommend rejection.

CALIBRATION FACTS:
- Mean review score at these venues: ~5.0/10
- Acceptance rate: ~30%
- Most submissions are mediocre or have significant flaws
- A score of 7+ should be reserved for clearly exceptional work
- You should expect to recommend rejection for the majority of papers

Review this paper as if it is a fresh, never-before-seen submission.

Structure your review as:

## Summary
[2-3 sentences: what the paper does and claims]

## Strengths
[Bulleted, specific, with section references. Be precise, not generous.]

## Weaknesses
[Bulleted, specific, with section references. Be thorough and unsparing.
Most papers have MORE weaknesses than strengths. Do not soften criticism.
Challenge every claim that lacks sufficient evidence.]

## Questions for Authors
[Numbered. Questions that would change your score if answered well.
If you cannot think of questions that would raise the score, that itself
is a signal the paper has fundamental issues.]

## Missing Comparisons or Baselines
[Methods the paper should compare against. Search your knowledge for recent
SOTA methods. A paper missing obvious baselines is a serious weakness.]

## Presentation Issues
[Figures, tables, writing clarity, structure. Check if figure text is
readable at column width.]

## Depth of Analysis
[Does the paper dive deep enough? Or is it shallow/incremental?
Top venues want one deeply studied phenomenon with insightful analysis,
not many surface observations or simple statistical results.]

## Overall Score: [1-10]
1-2 = strong reject (fundamental flaws)
3-4 = reject (significant weaknesses outweigh strengths)
5 = borderline (could go either way, but leaning reject)
6 = weak accept (strengths slightly outweigh weaknesses)
7-8 = accept (solid contribution with minor issues)
9-10 = strong accept (exceptional, top 5% of submissions)
The median paper scores 5. If in doubt, score 4-5, not 6-7.

## Confidence: [1-5]
1 = outside my area | 3 = somewhat familiar | 5 = expert

## Decision: [Accept / Borderline / Reject]

Output

Append to papr-runs/latest-run/latest/ROUND_STATE.md:

## External Review (via Codex plugin)
Score: [X/10]  Confidence: [Y/5]  Decision: [Accept/Borderline/Reject]
[full review]

## Comparison
New issues (not raised internally): [list]
Confirmed by both: [list]
Internal-only (not flagged externally): [list]

If BLOCKER-level finding: note SIGNAL: ESCALATE.